Applying a deep residual network coupling with transfer learning for recyclable waste sorting.

Recyclable waste sorting has become a key step for promoting the development of a circular economy with the gradual realization of carbon neutrality around the world. This study aims to develop an intelligent and efficient method for recyclable waste sorting by the method of deep learning. Thus, RWNet models, which refers to various ResNet structures (ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNet-152) based on transfer learning, were proposed to classify different types of recyclable waste. Cyclical learning rate and data augmentation were taken to improve the performance of RWNet models. In addition, accuracy, precision, recall, F1 score, and ROC were taken to evaluate the performance of RWNet models. Results showed that the accuracy of various RWNet models is almost at 88%, and the best accuracy is 88.8% in RWNet-152. The highest precision, recall, and F1 score in terms of weighted average value appeared in RWNet-101 (89.9%), RWNet-152 (88.8%), and RWNet-152 (88.9%), respectively. The area under the ROC curve (AUC) is higher than 0.9, except for the AUC value of plastic (0.85), which indicated that most of the recyclable waste can be well sorted by RWNet models. This study demonstrates the good performance of RWNet models that can be used to automatically sort most of the recyclable waste, which paves the way for better recyclable waste management.